Multi-domain No-reference Image Quality Assessment Based On Deep Learning

Image quality assessment is an important research branch in the field of computer vision,aiming at measuring image quality by subjective or objective means.Efficient quality evaluation algorithms can be applied to many visual fields,such as image inpainting,image synthesis,image super resolution,etc.After a long period of development,the target of image quality assessment has been extended from simple artificial distorted image to complex artificial distorted image and authentic distorted image,and the evaluation method has gradually changed from manual feature extraction to deep learning for end-toend evaluation.The increasing number of distortion types has brought new challenges to the field of image quality evaluation.At present,there is no general quality assessment algorithm to evaluate the quality of generated images in the field of text-to-image synthesis based on GAN.In order to fill this gap,thesis proposes a Multiple Metrics Quality Assessment for Birds dataset and an algorithm of no-reference generated image evaluation model.In the dataset,the excellent GAN model is used to obtain the generated images,and a set of evaluation criteria for the generated images conforming to human vision is proposed for manual annotation.Generalized Gaussian Distribution and Asymmetric Generalized Gaussian Distribution are used for feature extraction to obtain the pixel statistical features of the generated image.Then the Sparse Neighborhood Co-occurrence Matrix is established to obtain the generated image’s spatial information.The quality of the generated image is predicted by Support Vector Regression algorithm.The dataset proposed by us is the first to generated images from GAN,and it standardizes how to evaluate the generated images.The blind model created is also the first evaluation model for the single generated image.The PLCC and SRCC values of this evaluation model are 8.79% and 8.77% higher than those of the most advanced methods based on natural scene statistics,respectively.We also further discuss the rationality of the proposed special dataset and the application of the no-reference evaluation algorithm.The results show that the algorithm achieves convincing results both subjectively and objectively.One of the bottlenecks of image quality assessment is that the manual annotation of datasets is complicated and the label is difficult to obtain,which limits the scale of datasets.To solve this problem,we propose a semi-supervised generative adversarial regression network evaluation algorithm to evaluate different fields of distorted images without reference,including artificial distorted images and authentic distorted images.The algorithm extracts the features of labeled images,unlabeled images and generated images from the discriminator,so that the model can learn the similarity and authenticity between real images,and the difference between distorted images and generated image.Our model is trained with only 50% labeled images in the artificially distorted data set,and the evaluation effect is better than all traditional no-referenced evaluation algorithms and some supervised deep learning algorithms,with PLCC reaching 0.813 and SRCC reaching 0.827.The evaluation effect on authentic distorted image is also better than most evaluation algorithms.More importantly,the proposed evaluation algorithm relieves the dependence on manual annotation,which will be of great value for practical applications in related fields.